CN116863024A - Magnetic resonance image reconstruction method, system, electronic equipment and storage medium - Google Patents
Magnetic resonance image reconstruction method, system, electronic equipment and storage medium Download PDFInfo
- Publication number
- CN116863024A CN116863024A CN202310884048.2A CN202310884048A CN116863024A CN 116863024 A CN116863024 A CN 116863024A CN 202310884048 A CN202310884048 A CN 202310884048A CN 116863024 A CN116863024 A CN 116863024A
- Authority
- CN
- China
- Prior art keywords
- magnetic resonance
- convolution
- resonance image
- image reconstruction
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 63
- 239000011159 matrix material Substances 0.000 claims abstract description 27
- 238000010606 normalization Methods 0.000 claims description 12
- 238000012545 processing Methods 0.000 claims description 9
- 238000000605 extraction Methods 0.000 claims description 8
- 230000009466 transformation Effects 0.000 claims description 6
- 238000004590 computer program Methods 0.000 claims description 4
- 230000000694 effects Effects 0.000 abstract description 11
- 230000001133 acceleration Effects 0.000 description 13
- 238000012549 training Methods 0.000 description 12
- 230000035945 sensitivity Effects 0.000 description 11
- 238000002595 magnetic resonance imaging Methods 0.000 description 10
- 230000004927 fusion Effects 0.000 description 9
- 238000004364 calculation method Methods 0.000 description 7
- 238000010586 diagram Methods 0.000 description 7
- 238000005070 sampling Methods 0.000 description 7
- 238000012360 testing method Methods 0.000 description 4
- 238000013527 convolutional neural network Methods 0.000 description 3
- 230000010354 integration Effects 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 238000012795 verification Methods 0.000 description 3
- 238000004422 calculation algorithm Methods 0.000 description 2
- 238000007499 fusion processing Methods 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 238000011084 recovery Methods 0.000 description 2
- 230000001131 transforming effect Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 238000002059 diagnostic imaging Methods 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 210000000056 organ Anatomy 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000016776 visual perception Effects 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T11/00—2D [Two Dimensional] image generation
- G06T11/003—Reconstruction from projections, e.g. tomography
- G06T11/006—Inverse problem, transformation from projection-space into object-space, e.g. transform methods, back-projection, algebraic methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
- G06N3/0455—Auto-encoder networks; Encoder-decoder networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/30—Assessment of water resources
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Algebra (AREA)
- Mathematical Optimization (AREA)
- Mathematical Analysis (AREA)
- Pure & Applied Mathematics (AREA)
- Magnetic Resonance Imaging Apparatus (AREA)
Abstract
The invention discloses a magnetic resonance image reconstruction method, a system, electronic equipment and a storage medium, which relate to the field of image reconstruction, wherein the method comprises the steps of obtaining magnetic resonance original space data; generating undersampled data using the magnetic resonance raw spatial data and a mask matrix; reconstructing by using a rapid magnetic resonance image reconstruction network according to the undersampled data to obtain a reconstructed magnetic resonance image; two continuous convolution modules in each stage of the fast magnetic resonance image reconstruction network are an inverse Fourier convolution module and a Fourier convolution module which are connected in sequence. The invention can improve the reconstruction performance and effect of the image.
Description
Technical Field
The present invention relates to the field of image reconstruction, and in particular, to a magnetic resonance image reconstruction method, system, electronic device, and storage medium.
Background
Magnetic Resonance Imaging (MRI) is a method of non-invasively examining internal structures and organs of the body and has become an important diagnostic tool in medical imaging. MRI is performed by acquiring resonance signals of the nuclei in an object and converting these signals into a visualized image using computer techniques. However, acquisition of MRI images typically requires a long scanning process, which not only increases patient discomfort, but also increases cost and pressure on device usage.
To solve this problem, various methods of accelerating MRI have been proposed by researchers. Among them, the MRI-accelerated reconstruction from undersampled k-space data has become a popular method, which can shorten the scan time by reducing the number of sampling times, thereby improving the imaging speed. However, undersampling k-space data also introduces reconstruction problems such as artifacts and image blurring. Accordingly, researchers have proposed various methods to solve these problems.
Among methods of MRI-accelerated reconstruction from undersampled k-space data, compressed sensing (compressive sensing) is a common method that uses sparsity and low rank of the image for reconstruction. This method has been widely used in the reconstruction of MRI images and achieves good results.
In addition to compressed sensing, researchers have proposed some other methods to address the problem of MRI accelerated reconstruction from undersampled k-space data. For example, the multi-channel reconstruction method (multi-channel reconstruction) uses multiple independent channels for accelerated reconstruction, which can improve reconstruction quality and speed. Meanwhile, a Convolutional Neural Network (CNN) is also used in MRI image reconstruction, and can automatically learn the mapping from undersampled k-space data to a complete image, so that a better reconstruction effect can be obtained.
In general, MRI accelerated reconstruction from undersampled k-space data is an effective method that can greatly shorten scan times, thereby improving imaging speed and reducing patient discomfort. In practical applications, different methods and techniques are required to be combined, and selection and optimization are performed for specific problems to obtain the best reconstruction effect.
However, there are still many problems with these current algorithms. First, the reconstruction index of such methods is still not high enough, and for strict medical diagnostic requirements, near-perfect accurate, safe reconstruction results should be obtained. The magnetic resonance accelerated reconstruction still lacks sufficient accuracy, both from an index and visual perception point of view, and still requires a great deal of research effort to advance its reconstruction quality and accuracy.
Secondly, most of the methods are based on a depth convolution coding and decoding network to carry out optimal reconstruction on undersampled images, and the images after convolution and upsampling are subjected to problems of excessive smoothness of the images and difficult reconstruction due to detail loss, and limited receptive fields limit the reconstruction quality of the images, so that the application of the deep learning acceleration reconstruction technology in practical medical diagnosis is severely limited.
Disclosure of Invention
The invention aims to provide a magnetic resonance image reconstruction method, a magnetic resonance image reconstruction system, electronic equipment and a storage medium, which can improve image reconstruction performance and effect.
In order to achieve the above object, the present invention provides the following solutions:
a method of magnetic resonance image reconstruction, comprising:
acquiring magnetic resonance original space data;
generating undersampled data using the magnetic resonance raw spatial data and a mask matrix;
reconstructing by using a rapid magnetic resonance image reconstruction network according to the undersampled data to obtain a reconstructed magnetic resonance image; the rapid magnetic resonance image reconstruction network is a U-shaped network; two continuous convolution modules in each stage of the fast magnetic resonance image reconstruction network are an inverse Fourier convolution module and a Fourier convolution module which are connected in sequence.
Optionally, generating undersampled data by using the magnetic resonance raw spatial data and a mask matrix specifically includes:
and performing dot multiplication on the magnetic resonance original space data and the mask matrix to generate undersampled data.
Optionally, after generating undersampled data using the magnetic resonance raw spatial data and a mask matrix, further comprising:
and carrying out zero-mean normalization processing on the undersampled data.
Optionally, reconstructing by using a fast magnetic resonance image reconstruction network according to the undersampled data to obtain a reconstructed magnetic resonance image, which specifically includes:
the undersampled data passes through a first convolution layer of a fast magnetic resonance image reconstruction network to obtain convolution output;
performing feature extraction on the convolution output by using a convolution layer with residual errors of a rapid magnetic resonance image reconstruction network to obtain space adjacent feature information and multi-coil redundancy information;
carrying out channel adjustment on the spatial adjacent characteristic information and the multi-coil redundant information by utilizing a second convolution layer of a fast magnetic resonance image reconstruction network to obtain characteristic information;
utilizing an inverse Fourier convolution module of a fast magnetic resonance image reconstruction network according to the characteristic information to obtain image domain characteristics;
the image domain features are reconstructed by a third convolution layer of a fast magnetic resonance image reconstruction network to obtain a first cross-domain convolution result;
the cross-domain convolution result is utilized to reconstruct a Fourier convolution module of a network by utilizing a fast magnetic resonance image, so as to obtain a second cross-domain convolution result;
and splicing the first cross-domain convolution result and the second cross-domain convolution result to obtain a reconstructed magnetic resonance image.
The invention also provides a magnetic resonance image reconstruction system, comprising:
the acquisition module is used for acquiring magnetic resonance original space data;
the generation module is used for generating undersampled data by utilizing the magnetic resonance original space data and the mask matrix;
the reconstruction module is used for reconstructing by utilizing a rapid magnetic resonance image reconstruction network according to the undersampled data to obtain a reconstructed magnetic resonance image; the rapid magnetic resonance image reconstruction network is a U-shaped network; two continuous convolution modules in each stage of the fast magnetic resonance image reconstruction network are an inverse Fourier convolution module and a Fourier convolution module which are connected in sequence.
Optionally, the generating module specifically includes:
and the generating unit is used for carrying out dot multiplication on the magnetic resonance original space data and the mask matrix to generate undersampled data.
Optionally, the method further comprises:
and the zero mean normalization processing module is used for carrying out zero mean normalization processing on the undersampled data.
Optionally, the reconstruction module specifically includes:
the first convolution module is used for enabling the undersampled data to pass through a first convolution layer of a rapid magnetic resonance image reconstruction network to obtain convolution output;
the characteristic extraction unit is used for extracting the characteristics of the convolution output by utilizing a convolution layer with residual errors of a rapid magnetic resonance image reconstruction network to obtain space adjacent characteristic information and multi-coil redundancy information;
the channel adjusting unit is used for carrying out channel adjustment on the spatial adjacent characteristic information and the multi-coil redundant information by utilizing a second convolution layer of the rapid magnetic resonance image reconstruction network to obtain characteristic information;
the Fourier inverse transformation unit is used for obtaining image domain characteristics by utilizing an inverse Fourier convolution module of the fast magnetic resonance image reconstruction network according to the characteristic information;
the second convolution unit is used for reconstructing the image domain features by using a third convolution layer of the fast magnetic resonance image reconstruction network to obtain a first cross-domain convolution result;
the Fourier transform unit is used for obtaining a second cross-domain convolution result by utilizing a Fourier convolution module of the fast magnetic resonance image reconstruction network;
and the splicing unit is used for splicing the first cross-domain convolution result and the second cross-domain convolution result to obtain a reconstructed magnetic resonance image.
The present invention also provides an electronic device including:
one or more processors;
a storage device having one or more programs stored thereon;
the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the methods as described.
The invention also provides a computer storage medium having stored thereon a computer program which, when executed by a processor, implements a method as described.
According to the specific embodiment provided by the invention, the invention discloses the following technical effects:
the invention acquires magnetic resonance original space data; generating undersampled data using the magnetic resonance raw spatial data and a mask matrix; reconstructing by using a rapid magnetic resonance image reconstruction network according to the undersampled data to obtain a reconstructed magnetic resonance image; the rapid magnetic resonance image reconstruction network is a U-shaped network; two continuous convolution modules in each stage of the fast magnetic resonance image reconstruction network are an inverse Fourier convolution module and a Fourier convolution module which are connected in sequence. According to the invention, through the anti-Fourier convolution module and the Fourier convolution module which are sequentially connected in the rapid magnetic resonance image reconstruction network, the image reconstruction performance and effect are improved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are needed in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a schematic diagram of a magnetic resonance image reconstruction method provided by the invention;
FIG. 2 is a schematic diagram of an inverse Fourier convolution module;
FIG. 3 is a schematic diagram of a fast magnetic resonance image reconstruction network;
FIG. 4 is a schematic diagram of a reconstruction method based on an inverse Fourier convolution module and a Fourier convolution module;
fig. 5 is a diagram of a reconstruction result of the magnetic resonance image reconstruction method provided by the present invention;
FIG. 6 is a graph of the reconstruction result of an end-to-end network division method;
FIG. 7 is a graph of magnetic resonance image reconstruction results;
fig. 8 is a flowchart of a magnetic resonance image reconstruction method provided by the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The invention aims to provide a magnetic resonance image reconstruction method, a magnetic resonance image reconstruction system, electronic equipment and a storage medium, which can improve image reconstruction performance and effect.
In order that the above-recited objects, features and advantages of the present invention will become more readily apparent, a more particular description of the invention will be rendered by reference to the appended drawings and appended detailed description.
As shown in fig. 1 and 8, the magnetic resonance image reconstruction method provided by the present invention includes:
step 101: magnetic resonance raw spatial data is acquired.
Collecting and storing a large amount of magnetic resonance original k-space data as a model training, verifying and testing data set; when the training set and the verification set are acquired, full sampling scanning is carried out to obtain complete k-space original data, and the full sampling k-space data or a spatial domain form thereof can be used as a training truth value of a network for supervising the training of the network; the directly acquired data format is a k-space frequency domain complex format, the acquired frequency domain data is generally converted into an image domain by utilizing inverse Fourier transform, then the absolute value of the complex data is calculated, and then the square root sum is calculated in the multi-coil channel dimension, so that a final label image is obtained as supervision of a space domain network training stage; the acquisition of the test set can directly utilize the acceleration track of the device to directly acquire undersampled original k-space data, and store the data and the acceleration mask matrix.
Step 102: undersampled data is generated using the magnetic resonance raw spatial data and a mask matrix. Undersampled data is generated using the simulated acceleration mask matrix.
Step 102, specifically includes: and performing dot multiplication on the magnetic resonance original space data and the mask matrix to generate undersampled data.
In the model training and verification stage, generating and simulating a mask matrix in the undersampling process of equipment, wherein the generated mask matrix is used for designating phase encoding lines to be acquired, the mask matrix values of the positions are one, and the values of data points which are not acquired are zero; for two-dimensional data, continuous full sampling is performed on phase encoding lines in a certain range in the middle of all slices to serve as an automatic calibration area (for example, an area with 4 times of acceleration being 8% in the middle, an area with 8 times of acceleration being 4% in the middle and an area with 16 times of acceleration being 2% in the middle), and the rest positions are selected according to the acceleration rate and the sampling mode (equal interval or random sampling); alternatively, gaussian or spiral undersampled masks may be used. The mask matrix determines which positions of the magnetic resonance apparatus are scanned during data acquisition, and is automatically generated by a computer. The generation process is based on a logic that generally adopts a full acquisition strategy in the low frequency region: a region of a certain range of low frequency performs continuous full acquisition, for example, a region with 4 times acceleration of 8% in the middle of phase encoding, a region with 8 times acceleration of 4% in the middle, and a region with 16 times acceleration of 2% in the middle; the rest area is collected according to the adopted strategy, for example, 4 times of random sampling is adopted, the middle 8% of data is collected completely, 17% of phase encoding lines are collected randomly at the rest positions, and thus 25% of data are collected together, namely, the collection speed is accelerated by 4 times. After the mask matrix is obtained, the mask matrix is multiplied by the original k-space data point to simulate the accelerated scanning process of the magnetic resonance equipment, and undersampled data is generated.
After generating undersampled data using the magnetic resonance raw spatial data and a mask matrix, further comprising: and carrying out zero-mean normalization processing on the undersampled data.
After all data set data are obtained, zero-mean normalization is carried out on undersampled data before the undersampled data are sent to a network for training calculation, the mean value and the variance of all different channels of input data are calculated first, and then the normalized preprocessing of the input frequency domain and spatial domain data is carried out by utilizing the calculated mean value and variance, so that the training and convergence difficulty of the network can be reduced. In addition, after the network finishes forward calculation to obtain output, the output data is subjected to normalization removal operation by utilizing the obtained mean value and variance, and the output value is mapped back to the original value range so as to perform data consistency operation and reconstruction image index calculation. The standardized pretreatment specifically comprises the steps of firstly, respectively calculating the mean value and the variance of each channel; the data is then subtracted from the mean and then divided by the variance. De-normalization is the process of multiplying the processed data by the variance and then adding the mean.
Step 103: reconstructing by using a rapid magnetic resonance image reconstruction network according to the undersampled data to obtain a reconstructed magnetic resonance image; the rapid magnetic resonance image reconstruction network is a U-shaped network; two continuous convolution modules in each stage of the fast magnetic resonance image reconstruction network are an inverse Fourier convolution module and a Fourier convolution module which are connected in sequence. A fast magnetic resonance image reconstruction network based on double-domain ultra-fast Fourier convolution is designed, and the designed reconstruction network is trained by utilizing a training set and a verification set, so that the reconstruction network has the capability of high-quality reconstruction from undersampled data. The trained reconstruction model algorithm is embedded into the magnetic resonance equipment, so that the undersampled data acquired by the magnetic resonance equipment in an accelerating way are reconstructed in real time with high quality, and a display device is utilized to display a reconstructed image.
Step 103, specifically includes:
the undersampled data passes through a first convolution layer of a fast magnetic resonance image reconstruction network to obtain convolution output; performing feature extraction on the convolution output by using a convolution layer with residual errors of a rapid magnetic resonance image reconstruction network to obtain space adjacent feature information and multi-coil redundancy information; carrying out channel adjustment on the spatial adjacent characteristic information and the multi-coil redundant information by utilizing a second convolution layer of a fast magnetic resonance image reconstruction network to obtain characteristic information; utilizing an inverse Fourier convolution module of a fast magnetic resonance image reconstruction network according to the characteristic information to obtain image domain characteristics; the image domain features are reconstructed by a third convolution layer of a fast magnetic resonance image reconstruction network to obtain a first cross-domain convolution result; the cross-domain convolution result is utilized to reconstruct a Fourier convolution module of a network by utilizing a fast magnetic resonance image, so as to obtain a second cross-domain convolution result; and splicing the first cross-domain convolution result and the second cross-domain convolution result to obtain a reconstructed magnetic resonance image.
A newly designed fast Fourier deconvolution module is adopted in the fast magnetic resonance image reconstruction network to extract, fill and calculate the frequency domain characteristics of the magnetic resonance data; the fast Fourier convolution and inverse Fourier convolution module is designed to replace two continuous convolution modules at each stage in the original U-shaped network, so that a higher-quality double-domain reconstruction result can be obtained. The input is processed undersampled data or a feature map, and the output is refined feature map or a reconstructed image.
The fast Fourier deconvolution module has the following specific structure: the undersampled K space data or frequency domain features input to the modules (the first module input of a general network is the undersampled data output in the step 3, the later module inputs are all intermediate feature graphs, the output of the other modules except the last module is the intermediate feature graphs generally), the input data has complex characteristics, and half channels are designated as real channels and the other half channels are designated as corresponding imaginary channels through preset; the input data firstly passes through a convolution layer with a core size of 1*1, the data channel is transformed to the output channel, and the channel information is integrated and output as x 1 The method comprises the steps of carrying out a first treatment on the surface of the Then x 1 Extracting K space adjacent characteristic information and multi-coil redundant information through a convolution layer with residual error and a kernel size of 3*3, and outputting the information as x 2 The method comprises the steps of carrying out a first treatment on the surface of the Data x thereafter 2 Passing through a layer of core againThe convolution layer with the size of 1*1 integrates channel information and makes channel adjustment for the next cross-domain convolution, thereby expanding the frequency domain receptive field to all frequencies to obtain the characteristic x 3 The method comprises the steps of carrying out a first treatment on the surface of the Data x thereafter 3 Through a preset real part and an imaginary part, the characteristics of the corresponding image domain are obtained by utilizing Fourier inverse transformation to the image domain, the characteristics are input into a convolution layer with a kernel size of 1*1 for extracting airspace characteristics, and the frequency domain receptive field is expanded to all frequencies; after cross-domain convolution, the features are returned to the frequency domain by utilizing Fourier transformation, and residual error operation is carried out on the whole cross-domain convolution operation, so that smoothness of information flow is ensured, and the output is x 4 . Input x, which is then convolved across domains 3 And output x 4 And performing splicing operation, and then sending the splicing operation to a convolution layer with the last core size of 1*1 for channel integration and selection, wherein the output of the channel integration and selection is the output of the whole ultra-fast Fourier deconvolution module.
The network adopts a structure similarity loss function during training; the output of the network is reconstructed K space data, which are real numbers and imaginary numbers of multi-coil complex data respectively, the final output firstly calculates amplitude values through absolute values, and then the square root and the fused multi-coil image are used as a single image, so that the final reconstructed image is obtained.
For the reconstruction framework, an ultra-fast inverse Fourier convolution module is largely utilized to replace two continuous convolution layers at each stage in the U-shaped network, so that an ultra-fast inverse Fourier convolution-based network is formed, and the network has fewer parameter quantities, but the K space reconstruction performance is greatly improved.
The two-dimensional reconstruction method is in a parallel connection and cascade connection mode of an ultra-fast Fourier convolution network and an ultra-fast inverse Fourier convolution network. For multi-coil data, firstly, calculating a sensitivity map of each channel image by using an ultrafast Fourier convolution network (the sensitivity map is used for fusing the multi-coil data into a single-channel image), and then, carrying out reconstruction optimization in a double-domain parallel manner by using the ultrafast Fourier convolution network and an inverse Fourier convolution network respectively: in the frequency domain branch, firstly, multi-channel k-space data are utilized to fill multi-coil data, an ultrafast inverse Fourier convolution network is utilized to fill the multi-coil data, then a sensitivity map calculated before is utilized to fuse into a single image, and the single image is waited to fuse with the reconstruction result of another image domain; firstly, transforming an image domain network into a single image by using a sensitivity map, and then completing image domain reconstruction by using an ultrafast Fourier convolution network; the double-domain reconstruction result fusion process occurs in a single image domain, and the fused result is returned to the multi-coil k-space data format by using the sensitivity map and is used as the input or final output of the next stage after softer data consistency operation. And then repeatedly performing operations of multi-channel fusion, reconstruction, multi-channel recovery and data consistency, and repeatedly cascading for a plurality of times. The input of the method is multichannel undersampled k-space data and corresponding undersampled images thereof, and finally, a single image with multiple channels is output after absolute value calculation and square root and fusion.
Firstly, test data are input into a network to observe the reconstruction effect, each magnetic resonance undersampled data is input into a trained reconstruction network, the reconstruction of the data can be completed through forward propagation calculation, and the high-quality reconstruction result can be obtained through visualization and index calculation by the method, so that the image is clear and has no artifact.
Finally, embedding the network model into magnetic resonance equipment, obtaining undersampled data by utilizing a preset undersampled track, and then completing real-time and rapid magnetic resonance high-quality reconstruction by utilizing a rapid magnetic resonance image reconstruction method based on double-domain ultra-fast Fourier convolution, wherein the high-quality reconstruction is completed on the basis of 4 times, 8 times and even 16 times of acceleration.
The specific structure of the proposed ultra-fast inverse Fourier convolution is shown in FIG. 2, the application effect of the proposed ultra-fast inverse Fourier convolution in a typical U-shaped network is shown in FIG. 3, and the new network is called an ultra-fast inverse Fourier convolution-based U-shaped network and is used for completing k-space or frequency domain reconstruction. The undersampled K space data or the frequency domain characteristics are input to the module, the input data has complex characteristics, and half of channels are designated as real channels and the other half of channels are designated as corresponding imaginary channels through preset; firstly, input data passes through a convolution layer with a core size of 1*1, and a data channel is transformed to an output channel and channel information is integrated; then extracting K space adjacent characteristic information and multi-coil redundant information through a convolution layer with residual error and a kernel size of 3*3; then the data is integrated with channel information through a convolution layer with the kernel size of 1*1, and channel adjustment is carried out for the next cross-domain convolution, so that the frequency domain receptive field is expanded to all frequencies; then, the data is subjected to inverse Fourier transform to an image domain through a preset real part and an imaginary part to obtain corresponding image domain features, the features are input into a convolution layer with a kernel size of 1*1 to extract airspace features, and a frequency domain receptive field is expanded to all frequencies; after cross-domain convolution, the features are returned to the frequency domain by utilizing Fourier transformation, and meanwhile, residual error operation is carried out on the whole cross-domain convolution operation, so that smoothness of information flow is ensured. And then performing splicing operation on the input and output of the cross-domain convolution, and then sending the input and output to a convolution layer with the last core size of 1*1 for channel integration and selection, wherein the output is the output of the whole ultra-fast inverse Fourier convolution module.
The ultra-fast inverse Fourier convolution is used to replace two convolutions in a U-shaped network or a ResNet and other networks, and is used to provide a receptive field capable of covering the full frequency for the frequency domain reconstruction of the network.
In addition, based on the ultrafast inverse Fourier convolution and the ultrafast Fourier convolution (the two convolutions are collectively called as ultrafast Fourier convolution), a novel two-domain ultrafast Fourier convolution fast magnetic resonance image reconstruction method is provided, and a specific frame diagram is shown in figure 4.
The two-dimensional reconstruction method is in a parallel connection and cascade connection mode of an ultra-fast Fourier convolution network and an ultra-fast inverse Fourier convolution network. For multi-coil data, firstly, calculating a sensitivity map of each channel image by using an ultrafast Fourier convolution network (the sensitivity map is used for fusing the multi-coil data into a single-channel image), and then, carrying out reconstruction optimization in a double-domain parallel manner by using the ultrafast Fourier convolution network and an inverse Fourier convolution network respectively: in the frequency domain branch, firstly, multi-channel k-space data are utilized to fill multi-coil data, an ultrafast inverse Fourier convolution network is utilized to fill the multi-coil data, then a sensitivity map calculated before is utilized to fuse into a single image, and the single image is waited to fuse with the reconstruction result of another image domain; firstly, transforming an image domain network into a single image by using a sensitivity map, and then completing image domain reconstruction by using an ultrafast Fourier convolution network; the double-domain reconstruction result fusion process occurs in a single image domain, and the fused result is returned to the multi-coil k-space data format by using the sensitivity map and is used as the input or final output of the next stage after softer data consistency operation. And then repeatedly performing operations of multi-channel fusion, reconstruction, multi-channel recovery and data consistency, and repeatedly cascading for a plurality of times. The input of the method is undersampled data, and finally, a single image with absolute value calculation, square root and fusion multichannel is output.
A specific block diagram of the sensitivity map estimation module and the newly proposed softer data consistency module is provided below in fig. 4.
The softer data consistency module is used for firstly splicing the reconstructed k-space and the original undersampled data in the channel dimension, then carrying out data fusion and extraction through a 3*3 convolution operation, and automatically acquiring a proper data consistency strategy through network learning. And then the network learning process is further simplified through a mask, reconstruction results are directly adopted for the points which are not collected, and the data are automatically balanced through network training for the collected points.
Fig. 5 shows the reconstruction effect graph after 6 secondary ties. It can be seen that the network has very good accelerated reconstruction performance at 4 times acceleration.
Firstly, the existing reconstruction methods all use a common convolution structure, and the reconstruction receptive field of the network is seriously insufficient; secondly, the double domain is not utilized or the utilization of the double domain data is unreasonable, and the fusion of the double domain data is problematic, so that the related reconstruction performance is further limited; finally, more effective data consistency operation is not adopted, and data discontinuity is caused by hard mixed k-space, so that a reconstruction result is influenced.
The comparison is made with the end-to-end variation network by using a comparison test as a comparison example, which adopts only the reconstruction of a single-domain convolutional neural network and adopts the common data consistency operation. As shown in fig. 6 and 7. It can be seen that the image reconstructed by the method has higher signal-to-noise ratio, richer details and better reconstruction quality.
The method designs a novel double-domain ultra-fast Fourier convolution fast magnetic resonance image reconstruction frame, utilizes a centrosymmetric double-domain reconstruction structure, is more suitable for the characteristics of double-domain data, and expands the double-domain reconstruction receptive field to a full image; in addition, the single image fusion strategy of the image domain is simple and effective; a new softer data consistency operation is also utilized. The method has the characteristics of high reconstruction performance and good effect, and still has good reconstruction performance under high multiplying power.
These gains are mainly derived from the following four points:
1. application of dual domain ultrafast fourier convolution. On the basis of adopting the ultra-fast Fourier convolution in the image domain, the ultra-fast inverse Fourier convolution is further proposed to complete the reconstruction of the large receptive field of the k space.
2. The single image fusion strategy of the image domain is simpler and more effective, the fusion strategy is more reasonable, and the training and the learning of the network are easier.
3. A new softer data consistency operation is provided, so that different data consistency strategies can be adopted by the network at different frequencies, and a larger degree of learning freedom is given to the network.
4. The invention provides a novel double-domain ultra-fast Fourier convolution fast magnetic resonance image reconstruction frame, which has a centrosymmetric double-domain structure, so that a k domain can better utilize multi-coil redundant information, and an image domain can directly extract image details in a single image. Thus comprehensively obtaining excellent multi-coil reconstruction performance.
The invention also provides a magnetic resonance image reconstruction system, comprising:
and the acquisition module is used for acquiring the magnetic resonance original spatial data.
And the generation module is used for generating undersampled data by utilizing the magnetic resonance original space data and the mask matrix.
The reconstruction module is used for reconstructing by utilizing a rapid magnetic resonance image reconstruction network according to the undersampled data to obtain a reconstructed magnetic resonance image; the rapid magnetic resonance image reconstruction network is a U-shaped network; two continuous convolution modules in each stage of the fast magnetic resonance image reconstruction network are an inverse Fourier convolution module and a Fourier convolution module which are connected in sequence.
As an alternative embodiment, the generating module specifically includes:
and the generating unit is used for carrying out dot multiplication on the magnetic resonance original space data and the mask matrix to generate undersampled data.
As an alternative embodiment, further comprising:
and the zero mean normalization processing module is used for carrying out zero mean normalization processing on the undersampled data.
As an alternative embodiment, the reconstruction module specifically includes:
and the first convolution module is used for enabling the undersampled data to pass through a first convolution layer of a rapid magnetic resonance image reconstruction network to obtain convolution output.
And the characteristic extraction unit is used for carrying out characteristic extraction on the convolution output by utilizing a convolution layer with residual errors of the rapid magnetic resonance image reconstruction network to obtain space adjacent characteristic information and multi-coil redundancy information.
And the channel adjusting unit is used for carrying out channel adjustment on the spatial adjacent characteristic information and the multi-coil redundant information by utilizing a second convolution layer of the rapid magnetic resonance image reconstruction network to obtain the characteristic information.
And the Fourier inverse transformation unit is used for obtaining image domain characteristics by utilizing an inverse Fourier convolution module of the fast magnetic resonance image reconstruction network according to the characteristic information.
And the second convolution unit is used for reconstructing the image domain features by using a third convolution layer of the fast magnetic resonance image reconstruction network to obtain a first cross-domain convolution result.
And the Fourier transform unit is used for obtaining a second cross-domain convolution result by utilizing a Fourier convolution module of the fast magnetic resonance image reconstruction network.
And the splicing unit is used for splicing the first cross-domain convolution result and the second cross-domain convolution result to obtain a reconstructed magnetic resonance image.
The present invention also provides an electronic device including: one or more processors; a storage device having one or more programs stored thereon; the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the methods as described.
The invention also provides a computer storage medium having stored thereon a computer program which, when executed by a processor, implements a method as described.
In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant points refer to the description of the method section.
The principles and embodiments of the present invention have been described herein with reference to specific examples, the description of which is intended only to assist in understanding the methods of the present invention and the core ideas thereof; also, it is within the scope of the present invention to be modified by those of ordinary skill in the art in light of the present teachings. In view of the foregoing, this description should not be construed as limiting the invention.
Claims (10)
1. A method of magnetic resonance image reconstruction, comprising:
acquiring magnetic resonance original space data;
generating undersampled data using the magnetic resonance raw spatial data and a mask matrix;
reconstructing by using a rapid magnetic resonance image reconstruction network according to the undersampled data to obtain a reconstructed magnetic resonance image; the rapid magnetic resonance image reconstruction network is a U-shaped network; two continuous convolution modules in each stage of the fast magnetic resonance image reconstruction network are an inverse Fourier convolution module and a Fourier convolution module which are connected in sequence.
2. The method of magnetic resonance image reconstruction according to claim 1, wherein generating undersampled data using the magnetic resonance raw spatial data and a mask matrix, in particular comprises:
and performing dot multiplication on the magnetic resonance original space data and the mask matrix to generate undersampled data.
3. The method of magnetic resonance image reconstruction according to claim 1, further comprising, after generating undersampled data using the magnetic resonance raw spatial data and a mask matrix:
and carrying out zero-mean normalization processing on the undersampled data.
4. The method for reconstructing a magnetic resonance image according to claim 1, wherein the reconstructing is performed by using a fast magnetic resonance image reconstruction network according to the undersampled data, so as to obtain a reconstructed magnetic resonance image, specifically comprising:
the undersampled data passes through a first convolution layer of a fast magnetic resonance image reconstruction network to obtain convolution output;
performing feature extraction on the convolution output by using a convolution layer with residual errors of a rapid magnetic resonance image reconstruction network to obtain space adjacent feature information and multi-coil redundancy information;
carrying out channel adjustment on the spatial adjacent characteristic information and the multi-coil redundant information by utilizing a second convolution layer of a fast magnetic resonance image reconstruction network to obtain characteristic information;
utilizing an inverse Fourier convolution module of a fast magnetic resonance image reconstruction network according to the characteristic information to obtain image domain characteristics;
the image domain features are reconstructed by a third convolution layer of a fast magnetic resonance image reconstruction network to obtain a first cross-domain convolution result;
the cross-domain convolution result is utilized to reconstruct a Fourier convolution module of a network by utilizing a fast magnetic resonance image, so as to obtain a second cross-domain convolution result;
and splicing the first cross-domain convolution result and the second cross-domain convolution result to obtain a reconstructed magnetic resonance image.
5. A magnetic resonance image reconstruction system, comprising:
the acquisition module is used for acquiring magnetic resonance original space data;
the generation module is used for generating undersampled data by utilizing the magnetic resonance original space data and the mask matrix;
the reconstruction module is used for reconstructing by utilizing a rapid magnetic resonance image reconstruction network according to the undersampled data to obtain a reconstructed magnetic resonance image; the rapid magnetic resonance image reconstruction network is a U-shaped network; two continuous convolution modules in each stage of the fast magnetic resonance image reconstruction network are an inverse Fourier convolution module and a Fourier convolution module which are connected in sequence.
6. The magnetic resonance image reconstruction system according to claim 5, wherein the generation module specifically comprises:
and the generating unit is used for carrying out dot multiplication on the magnetic resonance original space data and the mask matrix to generate undersampled data.
7. The magnetic resonance image reconstruction system as set forth in claim 5, further comprising:
and the zero mean normalization processing module is used for carrying out zero mean normalization processing on the undersampled data.
8. The magnetic resonance image reconstruction system according to claim 5, wherein the reconstruction module comprises:
the first convolution module is used for enabling the undersampled data to pass through a first convolution layer of a rapid magnetic resonance image reconstruction network to obtain convolution output;
the characteristic extraction unit is used for extracting the characteristics of the convolution output by utilizing a convolution layer with residual errors of a rapid magnetic resonance image reconstruction network to obtain space adjacent characteristic information and multi-coil redundancy information;
the channel adjusting unit is used for carrying out channel adjustment on the spatial adjacent characteristic information and the multi-coil redundant information by utilizing a second convolution layer of the rapid magnetic resonance image reconstruction network to obtain characteristic information;
the Fourier inverse transformation unit is used for obtaining image domain characteristics by utilizing an inverse Fourier convolution module of the fast magnetic resonance image reconstruction network according to the characteristic information;
the second convolution unit is used for reconstructing the image domain features by using a third convolution layer of the fast magnetic resonance image reconstruction network to obtain a first cross-domain convolution result;
the Fourier transform unit is used for obtaining a second cross-domain convolution result by utilizing a Fourier convolution module of the fast magnetic resonance image reconstruction network;
and the splicing unit is used for splicing the first cross-domain convolution result and the second cross-domain convolution result to obtain a reconstructed magnetic resonance image.
9. An electronic device, comprising:
one or more processors;
a storage device having one or more programs stored thereon;
the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method of any of claims 1-4.
10. A computer storage medium, having stored thereon a computer program, wherein the computer program, when executed by a processor, implements the method of any of claims 1 to 4.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310884048.2A CN116863024A (en) | 2023-07-18 | 2023-07-18 | Magnetic resonance image reconstruction method, system, electronic equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310884048.2A CN116863024A (en) | 2023-07-18 | 2023-07-18 | Magnetic resonance image reconstruction method, system, electronic equipment and storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116863024A true CN116863024A (en) | 2023-10-10 |
Family
ID=88228429
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310884048.2A Pending CN116863024A (en) | 2023-07-18 | 2023-07-18 | Magnetic resonance image reconstruction method, system, electronic equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116863024A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117557675A (en) * | 2024-01-12 | 2024-02-13 | 北京航空航天大学杭州创新研究院 | Deep learning MRI image acceleration reconstruction method and system |
-
2023
- 2023-07-18 CN CN202310884048.2A patent/CN116863024A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117557675A (en) * | 2024-01-12 | 2024-02-13 | 北京航空航天大学杭州创新研究院 | Deep learning MRI image acceleration reconstruction method and system |
| CN117557675B (en) * | 2024-01-12 | 2024-04-30 | 北京航空航天大学杭州创新研究院 | Deep learning MRI image acceleration reconstruction method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107610194B (en) | Magnetic resonance image super-resolution reconstruction method based on multi-scale fusion CNN | |
| CN113077527B (en) | Rapid magnetic resonance image reconstruction method based on undersampling | |
| CN108828481B (en) | Magnetic resonance reconstruction method based on deep learning and data consistency | |
| Wen et al. | Transform learning for magnetic resonance image reconstruction: From model-based learning to building neural networks | |
| CN111951344B (en) | Magnetic resonance image reconstruction method based on cascade parallel convolution network | |
| CN113379867B (en) | Nuclear magnetic resonance image reconstruction method based on joint optimization sampling matrix | |
| CN106485764A (en) | The quick exact reconstruction methods of MRI image | |
| CN111436936B (en) | CT image reconstruction method based on MRI | |
| CN112164122A (en) | Rapid CS-MRI reconstruction method for generating countermeasure network based on depth residual error | |
| CN116863024A (en) | Magnetic resonance image reconstruction method, system, electronic equipment and storage medium | |
| CN111784792A (en) | Rapid magnetic resonance reconstruction system based on double-domain convolution neural network and training method and application thereof | |
| CN114419176A (en) | Reconstruction system and method for propeller magnetic resonance imaging | |
| CN113096207B (en) | Rapid magnetic resonance imaging method and system based on deep learning and edge assistance | |
| CN117611601B (en) | Text-assisted semi-supervised 3D medical image segmentation method | |
| CN115601572A (en) | Ultrasonic phased array image optimization reconstruction method and system based on semi-supervised CycleGan network | |
| CN112213673B (en) | Dynamic magnetic resonance imaging method, device, reconstruction computer and magnetic resonance system | |
| CN116070401A (en) | High-dimensional magnetic resonance image reconstruction method based on transform domain tensor low-rank priori depth expansion network | |
| CN113052840B (en) | Processing method based on low signal-to-noise ratio PET image | |
| CN115294300A (en) | Multi-branch attention prior parameterized finite projection fast fluorescence tomography reconstruction method | |
| CN112150570B (en) | Compressed sensing magnetic resonance imaging method based on iterative p-threshold projection algorithm | |
| CN114913262A (en) | Nuclear magnetic resonance imaging method and system based on joint optimization of sampling mode and reconstruction algorithm | |
| CN113920211A (en) | Rapid magnetic sensitivity weighted imaging method based on deep learning | |
| CN118135048A (en) | Interpolation reconstruction method, system, equipment and medium for undersampled magnetic resonance data | |
| CN117576250B (en) | Rapid reconstruction method and system for prospective undersampled MRI Dixon data | |
| CN117808919A (en) | MRI undersampling reconstruction method for generating countermeasure network based on deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |